We created a variety of Digital Terrain Models (DTM) and Digital Surface Models (DSM). These included:

• DSMs – buildings, trees and other features
• ‘Bald earth’ DTMs with buildings removed
• DTMs with additional building data like heights and roof features, but not vegetation

Each of these were then resampled to different horizontal grid resolutions (2m, 3m, 4m and 5m for LiDAR) to understand resolution effects. Hydrological data was based on gauge readings from the University of Hull which showed over 110mm of sustained rainfall with rates of over 6mm/hour. This was estimated to be more than a 150-year return period based on the Centre for Ecology and Hydrology (CEH) flood estimation handbook.

Methodology

To enable binary classification of a building’s flood status, we extracted building-level information from modelled flood depth grids. This was done using GIS (geographic information system) analysis and comparison with observed flood data. We determined:

• Correct wet predictions: buildings correctly predicted to flood
• Correct dry predictions: buildings correctly predicted not to flood
• False wet predictions: buildings predicted to flood which didn’t flood
• False dry predictions: buildings predicted not to flood which did flood

We assessed the predictive ability of each simulation and penalised over and under-prediction of flooded buildings. We also used the Kappa statistic to identify those model simulations which could better predict flooded and non-flooded buildings (as opposed to those which were correct due to chance).

Key findings

Not surprisingly, DSMs performed less well than either of the DTM types. We discovered that including building data actually reduced the models’ effectiveness and that this was primarily due to the overprediction of flooding caused by artefacts.

Of the DTM approaches, simulations based on DTM plus buildings slightly underperformed ‘bald earth’ DTM. This was due to the effect that adding detailed building information has upon floodplain storage.